Paolo Brambilla1,2; 1University of Milan, 2The University of Texas Health Science Center at Houston
In this work, a relationship between functioning and neurocognitive evaluations has been investigated using machine learning. The analyses were carried out on the neuropsychological scores in 707 subjects with clinical high risk (CHR), recent onset psychosis (ROP), recent onset depres-sion (ROD) and healthy controls (HC). Subjects were not divided in classes but the prediction has been considered as dealing with a continuum in terms of global functioning. Subjects with incomplete data were excluded from the analysis. The measures of Global Functioning (GF) social and role, which quantify how effectively the person is functioning in their everyday life, were used to predict the difference in functioning at baseline (T0) and at 9 months follow up (T1) in respect to lifetime. Then, 38 features drawn from the PRONIA cognitive battery (PCB) on the basis of a priori knowledge were entered into a model that was evaluated with machine learning methods implemented with Neurominer. The non-informative features were pruned and the data were scaled. The model under-went 10x10 cycles of internal cross-validation (CV1) and 10x10 cycles of cross-validation against an outer portion of the sample (CV2) that, at each cycle, did not enter the CV1. The results showed that the model could reliably predict a drop in social and role functioning at T0 with respect to both life and past year (p<0.01), and GF social could be predicted (p=0.03).
